The money machine is still running green.

NVIDIA just dropped another record quarter. Blackwell is accelerating faster than the politics swirling around H20. Share buybacks are on the horizon.

But will the US Government's new stake in Intel change market dynamics?

Especially with Beijing tripling down on domestic alternatives.

I’m Ben Baldieri. Every week, I break down what’s moving in GPU compute, AI infrastructure, and the data centres that power it all.

Here’s what’s inside this week:

Let’s get into it.

The GPU Audio Companion Issue #60

Want the GPU breakdown without the reading? The Audio Companion does it for you, but only if you’re subscribed. If you can’t see it below, click here to fix that.

NVIDIA Prints $46.7B, Blackwell Leads the Way

NVIDIA just posted another record quarter.

$46.7B in revenue, up 56% year-on-year. Data centre sales hit $41.1B, driven by a 17% sequential jump in Blackwell revenue. No H20 sales to China were booked, but NVIDIA still cleared a $180M release of previously reserved H20 inventory and shipped ~$650M of unrestricted units outside China. Gross margins? 72%+, with operating income up 31% QoQ to $28.4B.

Looking ahead, Q3 guidance is $54B, meaning NVIDIA is set to grow bigger than Intel and AMD combined on a quarterly run-rate. The company also added another $60B to its share buyback authorisation, doubling down on shareholder returns.

Why this matters:

  • Blackwell is now the centre of gravity with sequential revenue growth showing the ramp is accelerating, even as the H20 is kicked around like a political football.

  • Mid-70% gross margin guidance shows NVIDIA’s pricing power is holding despite mounting competitive and geopolitical headwinds.

  • With Europe, the US, and Asia all lining up a litany of national-level Blackwell projects, Team Green continues to entrench itself as the supplier of AI infrastructure.

Click here to read the full announcement from NVIDIA.

Intel Warns of Risks in Gov’t Stake

Intel now has the US government as its largest shareholder.

The 10% equity stake, worth ~$11bn, comes from converting CHIPS Act grants into stock and tying future disbursements to defence contracts. Trump called it “a deal that cost us zero.” Intel, in an SEC filing, called it a liability.

Why?

Approximately 76% of Intel’s revenue is generated outside of the United States. 29% of that comes from China alone. Having Washington as a shareholder ties Intel directly to US trade policy. Tariffs, sanctions, and all. That could make foreign buyers nervous, particularly in Asia, where Intel still relies on volume sales.

Why this matters:

  • 22% of Intel’s overall revenue comes from China.

  • Beijing is already actively discouraging domestic companies from purchasing both NVIDIA and AMD’s China-specific hardware due to Washington’s proposed 15% revenue cut. An equity stake will likely be perceived as an even greater risk by the Chinese Government.

  • If Intel’s SEC filing assertion is correct, the bottom-line impact will likely be significant. Expect shareholder lawsuits, customer flight, and retaliatory restrictions to follow.

Click here to read the full story on Reuters.

Vultr is plugging its GPU cloud directly into Digital Realty’s global data centre platform.

The two have announced a strategic tie-up that makes Vultr’s NVIDIA B200 and AMD MI325X clusters available on PlatformDIGITAL, with connectivity orchestrated via ServiceFabric and the AI Private Exchange (AIPx). For enterprises, that means instant access to pre-validated GPU environments designed for training, inference, and agentic AI, all delivered inside compliance boundaries and close to existing data.

Initial deployments are already live in Atlanta, Dallas, London, and Singapore, with Frankfurt, Mumbai, Sydney, and Tokyo next in line.

Why this matters:

  • Vultr is aggressively targeting corporates who need low-latency AI infra with audit trails and sovereignty baked in.

  • Digital Realty’s footprint gives Vultr reach into the world’s densest data hubs, aligning GPU supply with where enterprise data actually sits.

  • Secure interconnection via AIPx reduces the friction of running workloads across cloud, colo, and on-prem, creating a real on-ramp for “Private AI.”

Click here to read the full announcement on Capacity Media.

Grok 2, the Attention Play

xAI just open-sourced Grok 2.

The model is now on Hugging Face under the “Grok 2 Community License.” That means you can download and run it (42 files, ~500GB, 8 GPUs required). But you can’t use it to train new foundation models or escape xAI’s acceptable use policy.

Still, it gets Grok 2 in front of, and trending amongst, the one audience that matters for credibility: the Hugging Face community.

That’s the point.

In a market where distribution is the bottleneck, model quality alone is insufficient. Attention is the workaround. Dropping it on Hugging Face, and then suing Apple and OpenAI for monopolistic practices, even though Community Notes promptly flagged that claim as false, is a pretty effective way to stay part of the conversation.

Why this matters:

  • Lawsuits are a great way to drive attention, even when the community knows the claims are false.

  • Hugging Face is a systemically important channel whereby attention drives model adoption.

  • The lawsuit and release, therefore, are less about code and more about headlines, shifting the perception that Grok is locked out of scale.

Click here to read the full story on Hugging Face.

China Triples Down on Domestic AI Chips

Beijing is preparing to triple domestic AI chip output in 2026.

Three new fabs, all geared towards Huawei processors, are set to begin production between late this year and next. Fully ramped, their combined capacity could eclipse SMIC’s current AI lines. SMIC itself plans to double 7nm output, with Huawei still its largest customer, while players like Cambricon, Biren, and MetaX will gain more allocation as Nvidia’s China footprint contracts.

At the same time, Chinese chipmakers are converging around a common goal:

Optimising domestic silicon for the domestic AI model ecosystem.

DeepSeek’s adoption of FP8, a format that boosts efficiency at the cost of precision, is being treated as a proof point. Huawei’s 910D, Cambricon’s 690, and next-gen CXMT HBM3 samples are all being built to serve that ecosystem, alongside smaller fabs rushing their own versions.

Why this matters:

  • Industrial and geopolitical independence are central to Beijing’s vision for China, especially in key industries.

  • AI is one such industry, and tuning chips to meet the needs of Chinese model developers creates a homegrown positive feedback loop.

  • Given what’s just happened to Intel, recent comments from Tencent stating they don’t “need more GPUs”, and the success China has had in the past with a fully vertically integrated industrial strategy for a specific industry, this could be a turning point for global semiconductor market dynamics.

Click here to read the full story in the FT.

Meta’s Superintelligence Lab Is Leaking

Meta’s “Superintelligence Lab” is already losing talent.

Three researchers, Avi Verma, Ethan Knight, and Rishabh Agarwal, have resigned just weeks after joining the new unit. Two went back to OpenAI; Agarwal has not yet disclosed his next move. On top of that, Chaya Nayak, Meta’s director of generative AI product, is also leaving for OpenAI.

The exits come despite Mark Zuckerberg dangling nine-figure pay packages to lure senior AI talent. Insiders point to bureaucratic sprawl and repeated reshuffles of Meta’s AI groups as reasons why the lab’s momentum is already under question.

Even Shengjia Zhao, now installed as MSL’s chief scientist, had reportedly tried to return to OpenAI before taking the formal role.

Yet Meta is still pushing ahead.

New AI chief Alexandr Wang (ex-Scale AI) has inked a collaboration deal with Midjourney to accelerate video generation for Facebook and Instagram. But the contrast with OpenAI is sharp: instead of bleeding staff, OpenAI is pulling them back.

Why this matters:

  • Meta’s “superintelligence” bet might be dead in the water if even billion-dollar packages aren’t enough to ensure its marquee hires stick around.

  • With so many of the now ex-hires returning to OpenAI, it’s increasingly clear who still sets the cultural and technical centre of gravity in frontier AI.

  • Deals with Midjourney hint at a pivot to consumer-facing media products, while the bigger AGI push looks increasingly shaky.

Click here to read the full story from Wired.

Texas’ 11GW AI Campus Takes Shape

The Amarillo “HyperGrid” AI campus is now officially heading for nuclear.

Following on from the late June announcement, Fermi America has teamed up with Westinghouse to finalise its NRC licence application for four AP1000 reactors at the 11GW campus. Each reactor produces ~1.1 GW, meaning nuclear will eventually sit alongside gas, solar, and wind at a site slated for 18 million square feet of data centres.

Early site work has already begun, with 1GW due online by the end of 2026.

Why this matters:

  • Nuclear is moving from PowerPoint slides to actual licensing applications in AI data centres, and if successful, Amarillo could set a precedent for pairing reactors directly with hyperscale AI campuses.

  • That’s a big if, as Vogtle’s AP1000s came online $17bn over budget and seven years late.

  • While Fermi insists that this time will be different, aggressive deployment timelines combined with relatively experimental technologies in new implementation scenarios, against a historical backdrop of overruns, may well dampen investor appetite for the astronomical funding needed for project success.

Click here to read the full story on DCD.

The Rundown

The old free-market story, where chips flowed to the highest bidder, is breaking down.

Intel has effectively become a state-backed entity, with Washington converting CHIPS Act grants into equity. That creates a strategic supplier in name, but also a lightning rod.

Any chip tied to US policy will look like a political risk in Asia, where Intel still books significant sales.

On the other side, China is moving to triple domestic AI chip output.

2026 will see domestic fabs dedicated to Huawei and aligned around the Chinese model ecosystem. That’s not just about catching up to NVIDIA. It’s about ensuring that Beijing's industrial future is never again tied to the next US export ban.

What comes next?

Something closer to bloc-based industrial strategy.

That means each superpower building its own self-contained supply loop. Capital, talent, fabs, power, and models all aligned to sovereign priorities. For buyers, that means the market for compute will be less about price discovery and more about political alignment. For suppliers, it means that distribution is no longer global.

It’s gated.

And for NVIDIA, still straddling both worlds, the question is how long they can sell into every bloc before Washington or Beijing decides neutrality is no longer an option.

See you next week.

Reply

or to participate